Predicting surgical intensive care unit readmission with machine learning model: Bi-center training and validation

Patients readmitted to the surgical intensive care unit (SICU) face a high risk of mortality and increased hospital costs. Identifying patients at risk of SICU readmission is crucial. This study aims to develop a machine-learning (ML) model to predict SICU readmission.

This is a retrospective study based on data collected from the electronic healthcare records of Chang Gung Memorial Hospital. The model development cohort included adult patients admitted to the SICU from July 2020 to December 2022 at the Kaohsiung branch, while the external validation cohort consisted of patients admitted to the SICU from January 2023 to August 2023 at the Linkou branch. Various ML models, including Logistic Regression (LR), Random Forest, Gradient Boosting (GB), Artificial Neural Networks, and Support Vector Machines, were compared to determine the best model.

Of the 982 patients in the development cohorts, 68 (6.9%) experienced SICU readmission. The GB model outperformed other methods, achieving an AUROC of 0.82 (95% CI: 0.70-0.93) in the internal validation cohort. Eleven features significantly influence SICU readmission, with the central venous catheter usage days, the pre-ICU stay duration, blood urea nitrogen, and carbapenem usage days ranking as the top four important factors. The GB model also surpasses three previously published traditional logistic regression methods in the external validation cohort, with AUROCs of 0.80 (95% CI: 0.73-0.86), 0.73 (95% CI: 0.63-0.83), 0.70 (95% CI: 0.60-0.79), and 0.65 (95% CI: 0.50-0.73), respectively.

Machine learning models offer greater accuracy and reliability compared to traditional regression methods when predicting SICU readmission.